Until recently, choosing a local business was work the customer did by hand. Search a phrase, open a few sites, read the reviews, search again with different words, and piece together a shortlist. Most of that work left a trace you could see: a search you appeared in, a click on your site, a row in a report you own.
More of it is now handed to an assistant. A 2026 survey of 1,002 US adults found that 45% had used an AI tool for a local business recommendation in the past year, up from 6% a year earlier, which puts AI third as a source of recommendations behind Google and Facebook. Among people aged 30 to 44, it was 64%. The customer types the whole question once, the assistant does the searching and the reading, and the answer comes back naming two or three businesses. If yours is one of them, you may get a call. If it isn’t, nothing reaches you: no click, no impression, and no record that a customer was ever weighing you against someone else.
This guide shows you how to test it. The method below is the one we’d use by hand, and it will show you which of your topics produce an answer and who gets named in it. What it can’t settle is whether that’s the usual answer. Assistants answer the same question differently from one day to the next, and sometimes from one hour to the next, so a single check is one sample. Finding out what customers typically see means asking the same questions repeatedly, at scale, and across all of the topics where you’d want to be showing up.
The short version
- Nearly half of US adults used an AI tool for a local business recommendation in 2026, up from 6% a year earlier, and an AI answer names only two or three businesses.
- A business can test its own visibility in AI answers by asking ChatGPT, Claude, Gemini, Perplexity and Google the questions its customers ask, phrased with the town named and the business name left out, from a temporary or signed-out chat.
- Every AI answer lands in one of three states: the business’s site is linked as a source, the business is named without a link, or nothing is cited at all.
- AI assistants answer the same question differently from one run to the next; a 2026 study found only a third to two fifths of cited sources matched between two runs of the same question.
- In search a business holds a position; in AI answers it holds a probability, and improving it takes the same questions asked repeatedly across every topic that matters.
- Search results and AI answers have to be measured on the same topics, because a topic where a business appears in search but not in AI answers points at a specific page.
Why it’s worth testing your visibility in AI answers
Being left out of an AI answer is a different kind of loss from sitting on page two. On page two the customer scrolled past you. In an AI answer, they never heard of you.
The loss concentrates at the moment that matters. On one business we monitor, checked in September, the questions people ask when they’re choosing who to hire cited its site in none of eighteen answers across three of the four assistants. A competitor was cited in fifteen. This is a business that already appears in Google for those same topics; that’s how we knew to ask about them. Search was working, but the AI answer to the same question was sending potential customers somewhere else.
Half of the business’s top twenty topics had it in at least one AI answer. The other half didn’t. Which topics fall on which side is the thing you can’t guess and can only check.
The figures in this post come from accounts DepthFinder monitors, described without identifying the business. They show what one business saw over one period; the pattern is what carries across.
What to check and where
The place to begin is Google itself, because AI Overviews are the AI answer most of your customers already see. Search the way a customer would, as a question rather than a keyword, and if an AI Overview appears at the top of the page, read it and then expand the sources beneath it. Being named in the text and being one of the linked sources are two different outcomes, and the difference matters later, so it’s worth noting which one you got.
The assistants come next. ChatGPT, Claude, Gemini and Perplexity all answer questions about local businesses, and they answer them differently from each other, so a result from one says little about the others. Before you ask anything, open a temporary chat or a signed-out window. ChatGPT in particular remembers what you’ve told it, and a check from your own account will lean toward you in a way a customer’s never would. Then phrase the question as a customer would: put your town in it, and leave your business name out, since a customer who already knew your name wouldn’t be asking.
Each answer you get will land in one of three states. In the first, your site is linked as a source. In the second, your business is named but nothing on your site is linked, which usually means the assistant is working from a directory or a review site rather than from you. In the third, nothing is linked at all, because the assistant answered from what it already knew and never ran a search, so there was no shortlist for you to be on. Across the businesses we monitor, ChatGPT answers most understanding-stage questions that third way: when someone asks about the problem itself rather than who can fix it, it tends to answer from memory and cite no one. Keep a record of which state each answer produced and who else was named, because those are the two things you’ll want to compare against next time.
Where this gets harder is getting the same information from the tools you already use. Since the end of August, Google’s own reporting tool, Search Console, has shown how often your pages appeared in AI Overviews and AI Mode, by page, country and date, but it shows the appearance and nothing else: no clicks, and no record of which question produced it. For ChatGPT, Claude, Gemini and Perplexity there is no report at all. What those assistants say about your business is visible only to the person who asked, which is why the test above is the only way to see it without a tool built for the purpose.
Searching from the mindset of your customer
The result of your testing depends almost entirely on the questions you choose to put to the assistants. Consider an obvious test case like asking an AI assistant “what’s the best physiotherapist in Hamilton?” You’ll either be named or you won’t, and that will feel like an answer. The trouble is that “best” means something different to each person asking it. To one it means the clinic that treats runners, to another the one that bills insurance directly, to a third the one that can explain why their knee hurts on stairs, and each of them will phrase the question around the thing they care about rather than around the word “best.” An assistant that names you for the generic question can still leave you out of every one of those. Whether you come up consistently across the whole range of topics your customers care about is the real question, and it’s a much harder one to answer than a single search.
The most reliable place to find those topics is the searches Google has already recorded for your site. Google Search Console is a free report from Google that lists the searches your website has appeared in, and if you don’t spend time here often, the searches themselves will look terse: a few words with no question mark, typed in a hurry.
Most Google searches read like a question with the grammar stripped out. Someone who typed “knee pain stairs” wanted to know why their knee hurts when they go downstairs. Someone who typed “physio referral ontario” wanted to know whether they need a doctor’s note before they can book. Someone who typed “physio hamilton direct billing” was choosing which clinic to call. Reading the searches back into the questions behind them gives you a list of what your customers actually want to know, in their words rather than yours. (Site traffic is down, but are leads down too? covers what that report can and can’t tell you about your business.)
The people who Googled those three different searches are at different distances from booking an appointment, and the distance changes what an assistant does with the question. What stage are your customers at when they find you? describes the three stages in detail: understanding the problem, evaluating the options, and deciding who to hire.
For this test, take the four or five topics your business is best known for (or that you hope to be best known for) and write a question at each customer stage for each one. For a physiotherapy clinic in Hamilton, one topic might produce these three:
- Why does my knee hurt when I go down stairs?
- Do I need a referral to see a physiotherapist in Ontario?
- Which physiotherapy clinics in Hamilton do direct billing?
Each stage tends to produce a different kind of answer. The understanding question will often be answered without anyone being named, because the assistant is explaining a symptom and has no reason to recommend a clinic. The evaluating question is where a clinic with a page on the subject can find itself cited alongside the provincial health sources. The deciding questions deserve most of your attention, for two reasons. They’re the ones a customer asks right before choosing someone, so being left out costs the most. And across the businesses we monitor, they’re also where a business is most often missing from the AI answer even when it shows up in Google for the same topic.
Showing up in AI answers is a matter of odds
A search engine returns a ranked list, and if you ask it the same thing tomorrow you’ll generally get the same list. An assistant works differently. Each time it’s asked, it rewrites the question into a handful of searches, runs them, reads some of what comes back, and composes an answer from whatever it read, and every one of those steps can go a different way on the next attempt. (How AI search works, and why it’s not that different walks through the mechanics.) The practical consequence is that the same question, asked a dozen times, can produce a dozen answers that name different businesses.
A study published in April 2026 measured this directly by putting identical questions to ChatGPT, Gemini, Google AI Mode and Perplexity up to ten times each within a single day. Between any two runs of the same question, only about a third to two fifths of the cited sources were the same. The businesses named in the text were somewhat steadier, at a third to a half. The authors concluded that a single run says close to nothing about how often a business is named, and that it takes seven or more runs of the same question before a reliable picture emerges.
Our own monitoring shows the same behavior. On one business we monitor, asking the same question twice within an hour changed whether the site was cited on one question in twenty. Between July and September, three of six questions we had been tracking were answered differently, and on two of them the citation was gone.
This changes what being visible in AI answers means. In search, you hold a position. In AI answers, you hold a probability: your business is named in some share of the times the question is asked, and the work is to push that share up. A single test tells you three real things, which is that being named is possible, who else is being named, and which of your topics produce an answer at all. What it can’t tell you is what that share is, or whether it was higher last month.
Search and AI answers have to be measured together
If being named is a probability, then improving it is a matter of finding out where the odds are low and working on those topics until they rise. That takes two things a one-off test can’t provide. The first is repetition: the same questions asked again and again, on every topic that matters to your business rather than the handful you had time for, so that the share of answers naming you is a measured number instead of a guess. The second is search results on the same topics, side by side. The assistants build their answers from the same web pages Google ranks, so a topic where you appear in search but not in AI answers is telling you something specific about the page, and a topic where you appear in neither is telling you something else. Without both measurements on the same topics, you can see that you’re missing but not why.
DepthFinder measures both. It tracks AI answers in Google AI Overviews, ChatGPT, Claude, Gemini and Perplexity on the same topics it uses for your search results, so the two can be read against each other. The questions it asks are written from your own searches, three for every topic, built from its most-searched queries, with your area named and your business name left out, the way a customer would ask. Google’s AI answers are checked weekly on the topics that carry your demand and the assistants monthly, with both checked again after you tell us a page has changed. Every answer is sorted into the three states described above, so a name mention without a link and an answer with nothing cited each show up as what they are.
Your Visibility score already reflects whether Google’s AI answers cite you. What ChatGPT, Claude, Gemini and Perplexity cite is shown beside it, along with the number of topics checked. After connecting your Search Console, Google’s AI answers appear within days and the assistants by your first monthly check.